Added documentation and error checking for fake quantization

bin/instrument.py

@@ -25,7 +25,7 @@
 # Everytime the contents of compileOption is changed in input.yaml
 # this script should be run to create new fi.exe and prof.exe
 
-import sys, os, shutil
+import sys, os, shutil, math
 import yaml
 import subprocess
 
@@ -319,13 +319,25 @@ def readCompileOption():
           if (str(cOpt["tracingPropagationOption"]["generateCDFG"]).lower() == "true"):
             options["genDotGraph"] = True
 
-  if 'fakeQuant' in cOpt and (cOpt['fakeQuant']['targetLayer'] == 'conv' or cOpt['fakeQuant']['targetLayer'] == 'matmul'):
+  if 'fakeQuant' in cOpt:
+    targetLayer = cOpt['fakeQuant']['targetLayer']
     minPercentileOutlierThreshold = cOpt['fakeQuant'].get('minPercentileOutlierThreshold', 0)
     maxPercentileOutlierThreshold = cOpt['fakeQuant'].get('maxPercentileOutlierThreshold', 100)
-    bitWidth = cOpt['fakeQuant'].get('bitWidth')
+    bitWidth = cOpt['fakeQuant'].get('bitWidth', 8)
+
+    assert isinstance(targetLayer, str), "TargetLayer must be of type string" 
+    assert targetLayer == "conv" or targetLayer == "matmul" , "TargetLayer can only be 'conv' and 'matmul'" 
+    assert isinstance(minPercentileOutlierThreshold, int), "Minimum Percentile Value should be of type int"
+    assert isinstance(maxPercentileOutlierThreshold, int), "Maximum Percentile Value should be of type int"
+    assert isinstance(bitWidth, int), "BitWidth should be of type int"
+    assert 0 < bitWidth, "BitWidth must be a integer greater than 0"
+    assert math.log2(bitWidth).is_integer(), "BitWidth must be a exponent of power 2"
+    assert 0 <= minPercentileOutlierThreshold and minPercentileOutlierThreshold <= 100, "Minimum Percentile Value for Percentile should be greater than or equal to 0"
+    assert 0 <= maxPercentileOutlierThreshold and maxPercentileOutlierThreshold <= 100, "Maximum Percentile Value for Percentile should be greater than or equal to 0 and lesser than or equal to 100"
+    assert minPercentileOutlierThreshold <= maxPercentileOutlierThreshold, "Minimum Percentile Value should be lesser than Maximum Percentile Value"
 
     global fakeQuant
-    fakeQuant = [cOpt['fakeQuant']['targetLayer'], minPercentileOutlierThreshold, maxPercentileOutlierThreshold, bitWidth]
+    fakeQuant = [targetLayer, minPercentileOutlierThreshold, maxPercentileOutlierThreshold, bitWidth]
 
 ################################################################################
 def _suffixOfIR():

docs/input_masterlist_ml.yaml

@@ -139,3 +139,25 @@ runOption:
 
 kernelOption:
     - forceRun
+
+# To use Fake Quantization within LLTFI, you need to specify the target layer that would be Quantized 
+# There is a parameter minPercentileOutlierThreshold and maxPercentileOutlierThreshold which eliminates outliers within the tensor input 
+# However, this parameter needs to be mentioned by the user and is not automatically set
+# When set, the values permitted within the input tensor must satisfy this condition 
+# (floating number at minPercentileOutlierThreshold) <= input tensor <= (floating number at maxPercentileOutlierThreshold)
+# Values that dont satisfy this condition are shifted to the (floating number at minPercentileOutlierThreshold) or (floating number at maxPercentileOutlierThreshold),
+# whichever is the nearest
+# One sample code is - 
+
+compileOption:
+    fakeQuant: 
+      targetLayer: conv
+      minPercentileOutlierThreshold : 10
+      maxPercentileOutlierThreshold : 90
+      bitWidth: 8
+
+
+# Fake Qunatization parameters comes within compileOption. In here, you need to mention the targer layer - conv or matmul (currently only supporting these)
+# minPercentileOutlierThreshold and maxPercentileOutlierThreshold are optional parameters. On not mentioning these parameters, they are autoamtically set to 
+# minPercentileOutlierThreshold = 0 and maxPercentileOutlierThreshold = 100
+# bitWidth needs to be mentioned and is the bit width of the Quantized Int numbers which needs to be a exponent of power 2

sample_programs/ml_sample_programs/vision_models/cnn-fmnist/input.yaml

@@ -10,6 +10,13 @@ compileOption:
     regSelMethod: regloc
     regloc: dstreg
 
+    # Please uncomment this if you need to use the Fake Quantization Feature within LLTFI
+    # fakeQuant: 
+    #   targetLayer: conv
+    #   minPercentileOutlierThreshold : 10
+    #   maxPercentileOutlierThreshold : 90
+    #   bitWidth: 16
+
     includeInjectionTrace:
         - forward